#pip install --quiet transformers
#pip install beautifulsoup4
#pip install selenium
#pip install matplotlib
#pip install Pillow

import warnings
warnings.filterwarnings('ignore')
from selenium import webdriver
from selenium.webdriver.chrome.options import Options
from bs4 import BeautifulSoup
import time

url = "https://www.shutterstock.com/featured-collections/archive?sort=recently-added"

# Set up headless Chrome
options = Options()
options.headless = True
options.add_argument("--disable-blink-features=AutomationControlled")
options.add_argument("--no-sandbox")
options.add_argument("--disable-dev-shm-usage")
options.add_argument("user-agent=Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/123.0.0.0 Safari/537.36")

driver = webdriver.Chrome(options=options)
driver.get(url)

# Wait for JavaScript to load
time.sleep(5)

# Parse HTML after rendering
soup = BeautifulSoup(driver.page_source, "html.parser")

collection_links = []
for a in soup.find_all("a", href=True):
    href = a["href"]
    if "/featured-collections/" in href:
        if href.startswith("http"):
            collection_links.append(href)
        else:
            collection_links.append("https://www.shutterstock.com" + href)

collection_links = list(set(collection_links))
print(len(collection_links))
print(collection_links[:10])

driver.quit()

37
['https://www.shutterstock.com/featured-collections/dusty-dreams-324305362', 'https://www.shutterstock.com/featured-collections/dark-neon-glow-324702124', "https://www.shutterstock.com/featured-collections/new-year's-resolutions-324944189", 'https://www.shutterstock.com/featured-collections/into-wild-326008269', 'https://www.shutterstock.com/featured-collections/genai-holiday-edition-326669426', 'https://www.shutterstock.com/featured-collections/viva-la-vacay-325994693', 'https://www.shutterstock.com/featured-collections/sustainable-food-326377577', 'https://www.shutterstock.com/featured-collections/3d-illustrations-324701979', 'https://www.shutterstock.com/featured-collections/dreaming-monet-324702169', 'https://www.shutterstock.com/featured-collections/unlimited-image-326090319']

from selenium import webdriver
from selenium.webdriver.chrome.options import Options
from bs4 import BeautifulSoup
import time
from tqdm import tqdm
import re

options = Options()
options.headless = True  # Use False to debug
options.add_argument("--disable-blink-features=AutomationControlled")
options.add_argument("--no-sandbox")
options.add_argument("--disable-dev-shm-usage")
options.add_argument("user-agent=Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/123.0.0.0 Safari/537.36")

# Start the driver once
driver = webdriver.Chrome(options=options)

links = []
links_seen = set()

for source_url in tqdm(collection_links):
    retries = 3
    while retries > 0:
        try:
            driver.get(source_url)
            time.sleep(3)  # wait for JS to load

            page_source = driver.page_source
            jpg_regex = r'https://[^"]+\.jpg'
            image_links = re.findall(jpg_regex, page_source)

            for image_link in image_links:
                if image_link not in links_seen:
                    links_seen.add(image_link)
                    links.append({"image_link": image_link, "source_link": source_url})

            break  # success, exit retry loop

        except Exception as e:
            print(f"Error fetching {source_url}: {e}")
            retries -= 1
            time.sleep(5)  # wait before retry
            if retries == 0:
                print(f"Failed {source_url} after 3 retries")

driver.quit()
print(f"\n✅ Collected {len(links)} unique image links")
print(links[:5])

100%|██████████████████████████████████████████████████████████████████████████████████| 37/37 [02:51<00:00,  4.63s/it]

✅ Collected 7222 unique image links
[{'image_link': 'https://images.ctfassets.net/hrltx12pl8hq/2Bu8ZEmhePuuSNjL5LWUBw/b79b13ccd0693f8b73bfd42ea0b55cd4/Dust-Dreams.jpg', 'source_link': 'https://www.shutterstock.com/featured-collections/dusty-dreams-324305362'}, {'image_link': 'https://image.shutterstock.com/image-photo/gourmet-lady-vintage-dress-frill-260nw-772290127.jpg', 'source_link': 'https://www.shutterstock.com/featured-collections/dusty-dreams-324305362'}, {'image_link': 'https://image.shutterstock.com/image-photo/valle-de-la-luna-moon-260nw-1803845665.jpg', 'source_link': 'https://www.shutterstock.com/featured-collections/dusty-dreams-324305362'}, {'image_link': 'https://image.shutterstock.com/image-photo/face-elderly-man-feeling-headache-260nw-2488548353.jpg', 'source_link': 'https://www.shutterstock.com/featured-collections/dusty-dreams-324305362'}, {'image_link': 'https://image.shutterstock.com/image-photo/woman-young-green-light-travel-260nw-2336347791.jpg', 'source_link': 'https://www.shutterstock.com/featured-collections/dusty-dreams-324305362'}]

import requests
from PIL import Image
from io import BytesIO
import matplotlib.pyplot as plt

for i in range(10):
    # Fetch the image from URL
    response = requests.get(links[i]['image_link'])
    img = Image.open(BytesIO(response.content))
    
    # Plot the image
    plt.figure(figsize=(6, 6))
    plt.imshow(img)
    plt.axis('off')
    plt.title(f"Source: {links[i]['source_link']}", fontsize=10)
    plt.show()

#pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
#pip install protobuf
#pip install sentencepiece

# Load model directly
from transformers import AutoProcessor, AutoModel

processor = AutoProcessor.from_pretrained("google/siglip-so400m-patch14-384")
model = AutoModel.from_pretrained("google/siglip-so400m-patch14-384").eval()  # CPU by default   .to('cuda')

Using a slow image processor as `use_fast` is unset and a slow processor was saved with this model. `use_fast=True` will be the default behavior in v4.52, even if the model was saved with a slow processor. This will result in minor differences in outputs. You'll still be able to use a slow processor with `use_fast=False`.

nsample = 2000

import torch
from tqdm import tqdm
from PIL import Image
from torch.utils.data import DataLoader
import requests

# Initializes a big tensor torch_image_embeddings on CPU to hold all image embeddings 
# (dimension 1152, which is SigLIP’s embedding size).
# Define batch size and initialize tensor for embeddings
batch_size = 32  # Start with 32, adjust if needed
torch_image_embeddings = torch.zeros((len(links[:nsample]), 1152)).to('cpu')
used_links = []

# Defines a process_batch function that downloads images from their URLs, 
# preprocesses them with the SigLIP processor, runs them through the model 
# on GPU, and moves the resulting embeddings back to CPU.
def process_batch(batch_links):
    images = []
    _used_links = []
    for l in batch_links['image_link']:
        if l in used_links:
            continue
        try:
            images.append(Image.open(requests.get(l, stream=True).raw))
            _used_links.append(l)
        except:
            print(f'error. Skipping {l}')

    with torch.no_grad():
        # Move only the input to the GPU
        #image_inputs = processor(images=images, return_tensors="pt").to("cuda")
        image_inputs = processor(images=images, return_tensors="pt")
        batch_embeddings = model.get_image_features(**image_inputs)
        batch_embeddings = batch_embeddings.cpu()  # Move back to CPU after processing

    return _used_links, batch_embeddings

# Uses a PyTorch DataLoader to split links into batches of 32.
data_loader = DataLoader(links[:nsample], batch_size=batch_size, shuffle=False)

# Iterates through the batches, computes embeddings, and fills them 
# into the pre-allocated tensor in order.
start_idx = 0
for batch_links in tqdm(data_loader, total=len(data_loader)):
    try:
        __used_links, batch_embeddings = process_batch(batch_links)
        end_idx = start_idx + batch_embeddings.shape[0]
        torch_image_embeddings[start_idx:end_idx] = batch_embeddings
        start_idx = end_idx
        
        # Keeps track of used_links to avoid duplicates,
        used_links += __used_links

        ## Free up GPU memory after each batch
        #del batch_embeddings
        #torch.cuda.empty_cache()

    except Exception as e:
        print(e)
        print(f'error. Skipping batch')

  3%|██▌                                                                             | 2/63 [07:03<3:35:23, 211.87s/it]

error. Skipping https://image.shutterstock.com/image-photo/floral-sunlight-shadows-on-neutral-260nw-2498094767.jpg
error. Skipping https://image.shutterstock.com/image-photo/floral-sunlight-shadows-on-neutral-600w-2498094767.jpg

  8%|██████▎                                                                         | 5/63 [17:57<3:32:23, 219.71s/it]

error. Skipping https://image.shutterstock.com/image-photo/golden-eagle-aquila-chrysaetos-one-260nw-2498069731.jpg
error. Skipping https://image.shutterstock.com/image-photo/golden-eagle-aquila-chrysaetos-one-600w-2498069731.jpg

 10%|███████▌                                                                        | 6/63 [21:14<3:21:25, 212.02s/it]

error. Skipping https://image.shutterstock.com/image-photo/horizontal-medium-closeup-portrait-african-260nw-1977500189.jpg
error. Skipping https://image.shutterstock.com/image-photo/horizontal-medium-closeup-portrait-african-600w-1977500189.jpg

 37%|████████████████████████████                                                 | 23/63 [1:22:49<2:21:59, 212.99s/it]

error. Skipping https://image.shutterstock.com/image-photo/group-emperor-penguins-stands-together-260nw-2561901117.jpg

 38%|█████████████████████████████▎                                               | 24/63 [1:26:14<2:16:55, 210.65s/it]

error. Skipping https://image.shutterstock.com/image-photo/group-emperor-penguins-stands-together-600w-2561901117.jpg

torch_image_embeddings = torch_image_embeddings[:len(used_links)]
torch_image_embeddings.shape

torch.Size([1984, 1152])

model.get_image_features(**processor(images=Image.open(requests.get(used_links[0], stream=True).raw), 
                                     return_tensors="pt"))[0][:5]

tensor([0.1646, 0.3025, 0.0543, 0.3380, 0.1341], grad_fn=<SliceBackward0>)

torch_image_embeddings.shape

torch.Size([1984, 1152])

torch_image_embeddings[0][:5]

tensor([0.1646, 0.3025, 0.0543, 0.3380, 0.1341])

import torch.nn.functional as F
from IPython.display import Image, display
# Define candidate labels

def generate_stock_photo_recs(stock_photo_requests, k = 5):
    """
    This code applies RAG for images
    stock_photo_requests: list of basic queries
    k: number of images to return
    """

    # Preprocess inputs
    #text_inputs = processor(text=stock_photo_requests, return_tensors="pt", padding=True).to('cuda')
    # pass the stock_photo_requests to processor which has a tokenizer, processor tokenize all of the queries
    text_inputs = processor(text=stock_photo_requests, return_tensors="pt", padding=True)
    
    # get text embedding from the processor ouyput text_inputs,
    torch_text_embeddings = model.get_text_features(text_inputs.input_ids).cpu()

    # Compute cosine similarity scores using torch.cosine
    cosine_scores = torch.cosine_similarity(torch_text_embeddings, torch_image_embeddings.unsqueeze(1), dim=-1)

    image_indices = cosine_scores.argsort(dim=0, descending=True).T.tolist()
    image_scores = cosine_scores.sort(dim=0, descending=True).values.T.tolist()
    probs = torch.sigmoid(torch.as_tensor(image_scores)).tolist()

    results = []
    for i, (image_index, image_score) in enumerate(zip(image_indices, image_scores)):
        top_k_images = image_index[:k]
        top_k_scores = image_score[:k]
        results.append({
            "request": stock_photo_requests[i],
            "top_k_images": [used_links[t] for t in top_k_images],
            "top_k_scores": top_k_scores
        })

    return results

import torch
import torch.nn.functional as F
from IPython.display import Image, display

def generate_stock_photo_recs(stock_photo_requests, k=5):
    """
    Retrieve relevant stock photo recommendations using text-to-image retrieval.

    Args:
        stock_photo_requests (list[str]): List of textual queries describing the desired stock photos.
        k (int, optional): Number of top images to return for each query. Default is 5.

    Returns:
        list[dict]: A list of results, where each item contains:
            - "request": The original query string.
            - "top_k_images": List of top-k image URLs or paths.
            - "top_k_scores": List of similarity scores for the top-k images.
    """

    # Tokenize text queries using the processor
    text_inputs = processor(
        text=stock_photo_requests,
        return_tensors="pt",
        padding=True
    )
    
    # Extract text embeddings from the model
    text_embeddings = model.get_text_features(
        text_inputs.input_ids
    ).cpu()

    # Compute cosine similarity between text and all image embeddings
    cosine_scores = torch.cosine_similarity(
        text_embeddings, 
        torch_image_embeddings.unsqueeze(1), 
        dim=-1
    )

    # Sort indices and scores in descending order
    sorted_indices = cosine_scores.argsort(dim=0, descending=True).T.tolist()
    sorted_scores = cosine_scores.sort(dim=0, descending=True).values.T.tolist()

    # Apply sigmoid normalization to scores
    normalized_scores = torch.sigmoid(torch.tensor(sorted_scores)).tolist()

    results = []
    for i, (indices, scores) in enumerate(zip(sorted_indices, sorted_scores)):
        top_k_indices = indices[:k]
        top_k_scores = scores[:k]
        results.append({
            "request": stock_photo_requests[i],
            "top_k_images": [used_links[idx] for idx in top_k_indices],
            "top_k_scores": top_k_scores
        })

    return results

stock_photo_requests = ["Beautiful scenary from airplane window at evening", "A person eating watermelon",]
results = generate_stock_photo_recs(stock_photo_requests, k=1)
results

[{'request': 'Beautiful scenary from airplane window at evening',
  'top_k_images': ['https://image.shutterstock.com/z/stock-photo-a-passanger-looking-out-the-airplane-window-during-sunset-2447397523.jpg'],
  'top_k_scores': [0.13592387735843658]},
 {'request': 'A person eating watermelon',
  'top_k_images': ['https://image.shutterstock.com/z/stock-photo-cute-little-girl-eating-watermelon-on-the-grass-in-summertime-54065323.jpg'],
  'top_k_scores': [0.12953577935695648]}]

from PIL import Image

for result in results:
    print(f'Top image for request "{result["request"]}":')
    for image_link in result['top_k_images']:
        # shrink to fit
        print(f'Score: {result["top_k_scores"][result["top_k_images"].index(image_link)]}')
        display(Image.open(requests.get(image_link, stream=True).raw).resize((256, 256)))

Top image for request "Beautiful scenary from airplane window at evening":
Score: 0.13592387735843658

Top image for request "A person eating watermelon":
Score: 0.12953577935695648

#pip install python-dotenv

import os
from dotenv import load_dotenv

load_dotenv(dotenv_path='.env', override=True)  # reads .env file
api_key = os.getenv("OPENAI_API_KEY")

print(api_key[:8])  # just to check it's loaded

sk-proj-

from openai import OpenAI
#from google.colab import userdata
from langchain_openai import ChatOpenAI


openai_client = OpenAI(
    api_key=api_key
)

# Initialize LLM
llm = ChatOpenAI(
    model="gpt-4o-mini",   # or "gpt-4o", "gpt-3.5-turbo", etc.
    temperature=0          # 0 = deterministic, higher = more creative
)

#pip install langchain_openai
#pip install langchain langchain-community -q

scene_description = 'women or children eating or preparing food'

from langchain.output_parsers import CommaSeparatedListOutputParser
from langchain.prompts import PromptTemplate

# Define parser
parser = CommaSeparatedListOutputParser()

# Build prompt
prompt = PromptTemplate(
    template="think through and write 5 example stock photo titles for a scene : {scene_description}. I need the examples as a list of strings. Only list, not extra information without numbering",
    input_variables=["k"],
    partial_variables={"format_instructions": parser.get_format_instructions()}
)

# Call the model
chain = prompt | llm | parser
stock_photo_requests = chain.invoke({"scene_description": scene_description})

print(stock_photo_requests)

['Joyful Children Baking Cookies Together in a Bright Kitchen  ', 'Mother and Daughter Preparing a Colorful Salad in the Garden  ', 'Smiling Friends Cooking Pasta and Sharing Laughter in the Kitchen  ', 'Little Chefs Decorating Cupcakes with Colorful Frosting  ', 'Women Enjoying a Picnic with Homemade Dishes in a Sunny Park  ']

results = generate_stock_photo_recs(stock_photo_requests, k=1)
results

[{'request': 'Joyful Children Baking Cookies Together in a Bright Kitchen  ',
  'top_k_images': ['https://image.shutterstock.com/z/stock-photo-three-generation-males-decorating-gingerbread-house-at-christmas-2221897031.jpg'],
  'top_k_scores': [0.10619477927684784]},
 {'request': 'Mother and Daughter Preparing a Colorful Salad in the Garden  ',
  'top_k_images': ['https://image.shutterstock.com/z/stock-photo-a-woman-joyfully-examines-green-tomatoes-in-her-garden-the-lush-plants-and-modern-cabin-backdrop-2503619849.jpg'],
  'top_k_scores': [0.14137211441993713]},
 {'request': 'Smiling Friends Cooking Pasta and Sharing Laughter in the Kitchen  ',
  'top_k_images': ['https://image.shutterstock.com/z/stock-photo-happy-multigenerational-people-having-fun-doing-barbecue-grill-at-house-rooftop-summer-gatherings-2298094479.jpg'],
  'top_k_scores': [0.10109487175941467]},
 {'request': 'Little Chefs Decorating Cupcakes with Colorful Frosting  ',
  'top_k_images': ['https://image.shutterstock.com/z/stock-photo-three-generation-males-decorating-gingerbread-house-at-christmas-2221897031.jpg'],
  'top_k_scores': [0.0807461068034172]},
 {'request': 'Women Enjoying a Picnic with Homemade Dishes in a Sunny Park  ',
  'top_k_images': ['https://image.shutterstock.com/image-photo/woman-preparing-summer-salad-fresh-600w-2622974005.jpg'],
  'top_k_scores': [0.14676518738269806]}]

from PIL import Image

for result in results:
    print(f'Top image for request "{result["request"]}":')
    for image_link in result['top_k_images']:
        # shrink to fit
        print(f'Score: {result["top_k_scores"][result["top_k_images"].index(image_link)]}')
        display(Image.open(requests.get(image_link, stream=True).raw).resize((256, 256)))

Top image for request "Joyful Children Baking Cookies Together in a Bright Kitchen  ":
Score: 0.10619477927684784

Top image for request "Mother and Daughter Preparing a Colorful Salad in the Garden  ":
Score: 0.14137211441993713

Top image for request "Smiling Friends Cooking Pasta and Sharing Laughter in the Kitchen  ":
Score: 0.10109487175941467

Top image for request "Little Chefs Decorating Cupcakes with Colorful Frosting  ":
Score: 0.0807461068034172

Top image for request "Women Enjoying a Picnic with Homemade Dishes in a Sunny Park  ":
Score: 0.14676518738269806

Scrape shutterstock ¶

Download Siglip (Sigmoid Loss for Language-Image Pretraining)¶

Batch Processing Scraped Images into SigLIP Embeddings¶

Prompt-Driven Stock Image Title Generation¶

Limitations of CLIP¶

Scrape shutterstock¶

Download Siglip (Sigmoid Loss for Language-Image Pretraining)¶

Batch Processing Scraped Images into SigLIP Embeddings¶

Prompt-Driven Stock Image Title Generation¶

Limitations of CLIP¶

Scrape shutterstock ¶